Why does generating a transcript have to be the end of the workflow?
Transcription Portal started as a simple internal tool for converting audio and video into text. By the time I worked on v2.0, the problem had shifted. Users were spending more time reviewing, correcting, searching and working with transcripts than generating them.
My role was to redesign the experience around that shift, turning a one-time transcription flow into a persistent workspace where generated content could be reviewed, edited and explored in context.
Role
Product Designer
Product
PoC Transcription Portal
Project type
AI-assisted Product, Workflow & Workspace Redesign
Primary focus
Workflow + Interaction Design
Scope
v2.0 Product Redesign
Team
Product · Design · Engineering
Key areas
Transcript Workspace
AI Generation
Editing
Search
Timestamps
Media Context
Processing States
Overview
A transcription tool that had become a workspace
The first version of Transcription Portal followed a simple model: upload a file, wait for transcription, and download the result.
As the product evolved, that model became limiting. Longer recordings meant longer transcripts. Users needed to correct generated content, locate specific moments and understand the transcript against the original media.
The product therefore needed to support the work after transcription, not just the generation itself.
The assignment was to redesign the workflow, not the transcription model
The transcription engine was an existing part of the product. The redesign focused on how people interacted with its output.
That meant rethinking the workspace around the transcript itself: how generation states were communicated, how users moved between media and text, how generated content could be edited, and how long transcripts could be searched and navigated.
The goal was not to make AI feel more autonomous. It was to make its output more usable and controllable.
WHAT CHANGED
The product had outgrown the upload-and-download model
The first version of Transcription Portal was built around a simple task:
Upload a file → wait for transcription → download the result.
That model worked when transcription was the destination. But as recordings became longer and users spent more time reviewing the generated text, the transcript itself became the place where the work happened.
V2 therefore shifted the product from a one-time generation flow toward a persistent workspace where users could return to transcripts, review content, make corrections and find specific moments without leaving the product.
The shift was from generating text to working with it
The redesign was not about adding more AI features. It was about giving the output of the existing system a useful place to live.
The product now had to support what happened after generation
The workflow expanded from:
That change became the foundation for the rest of the design.
THE PRODUCT PROBLEM
Generating the transcript was only the first problem
Once a transcript existed, users still had to make sense of it.
Long-form transcripts were difficult to scan. Small transcription errors could require disproportionate correction effort. Finding a specific statement meant moving through large amounts of text, while the relationship between the transcript and its original audio or video needed to remain clear.
The product was producing text successfully, but the workflow around that text was creating the friction.
The transcript needed context
The workflow expanded from:
A transcript could not be treated as an isolated document.
Users needed to move between:
MEDIA → TIMESTAMP → TRANSCRIPT → EDIT
while retaining enough context to understand where a piece of content came from.
That made the transcript workspace more than a reading surface. It became the primary interface for reviewing and working with generated content.
The harder problem was control
AI-generated text could not be treated as authoritative.
The system needed to communicate when transcription was processing, when it had completed, and when something had failed. Once generated, the content also needed to remain editable so users could correct or restructure it themselves.
That created a simple product principle:
AI generates the content. The user decides what stays.
This principle shaped the workspace, editing interactions, processing states and search experience.
MY ROLE
The assignment was to redesign the workflow around an existing transcription capability
The transcription engine itself was not the design problem.
My responsibility was to rethink how users moved through the product once a file entered the system, from generation and processing states through transcript review, editing and search.
I owned the experience across the core workflow, including the structure of the workspace, transcript interaction, navigation, states and supporting interactions.
The work had to balance speed with control
The product needed to feel fast and straightforward, but reducing the workflow to a single AI-generated result would remove the controls users needed once the output was imperfect.
So the design direction became less about adding layers around AI and more about making the existing workflow predictable, editable and easy to navigate.
The central question was:
How do you make AI-generated content useful without making the user work around the AI?
That became the basis for the design decisions that followed.
DESIGN WORK
The interface had to support the work after AI generated the transcript
The redesign was not one large interaction problem. It was a set of smaller decisions around how generated content should behave once it entered the product.
I focused on four areas where the workflow could either become useful or fall apart.
01. Keep the transcript connected to its source
A transcript without context is difficult to verify
Instead of treating the generated text as a standalone document, the workspace brought together:
Media → Timestamps → Transcript → Search → Editing
This allowed users to move between what was said and where it occurred in the original recording.
The transcript is not the source. It is an interpretation of the source.
02. Make generation states visible
AI introduced states the interface could not hide
Transcription was not instantaneous. The product needed to communicate what was happening while the system processed the file.
The dashboard therefore accounted for distinct states:
In Progress → Generated → Failed
Each state needed to answer a basic question:
What is happening to my file right now?
Rather than treating loading as a temporary UI detail, I treated processing as part of the product workflow.
03. Keep AI output editable
Generated did not mean final
A transcription model can produce useful output without producing perfect output.
The workspace therefore treated the transcript as something the user could modify rather than a locked AI result.
Users could:
Edit → Add Line → Merge Line → Correct
This shifted the relationship between the user and the model.
The AI produced the first version.
The user remained responsible for the final content.
The important interaction was not "Accept AI."
It was:
Give the user enough control to correct the AI without having to leave the workflow.
04. Make long-form content searchable
Finding something should not require reading everything
As transcripts became longer, navigation became a product problem of its own.
Search allowed users to move directly from a query to relevant content instead of manually scanning the transcript.
But search also had to preserve context.
The useful result was not simply:
"Here is the matching text."
It was:
"Here is the matching text, and here is where it exists within the transcript and source."
KEY DECISIONS
The design was shaped by four decisions
Rather than presenting generic design principles, I would make this section a sequence of Decision → Why → Trade-off → Result.
That makes your judgement visible.
Keep the transcript attached to the media
DECISION
Treat the transcript and source media as one working context.
WHY
Generated text can lose meaning when separated from the recording it represents.
TRADE-OFF
A richer workspace is more complex than a document-only interface.
RESULT
Users can move between transcript content, timestamps and the original recording without leaving the primary workflow.
Treat AI output as editable
DECISION
Make correction part of the core transcript experience.
WHY
Generated content can contain errors, and users still need authority over the final result.
TRADE-OFF
Editing introduces additional interaction and state complexity.
RESULT
The transcript becomes a working document rather than a generated artifact.
Make processing state explicit
DECISION
Represent generation as a visible product state.
WHY
Users need to understand whether their file is processing, ready or unsuccessful.
TRADE-OFF
More states mean more UI and more edge cases to design.
RESULT
The system communicates its status instead of leaving users to infer what happened.
Search the content, not just the file
DECISION
Make transcript content directly searchable.
WHY
Once a recording has been transcribed, the generated text becomes a new way of navigating the original content.
TRADE-OFF
Search requires the transcript to behave like structured content rather than a static output file.
RESULT
Users can locate information inside long-form recordings without manually scanning the entire transcript.
REFLECTION
AI output is not the product. The workflow around it is.
Designing Transcription Portal changed how I approached AI-generated content.
A model can produce an answer quickly, but that does not make the answer usable. People still need to understand where it came from, verify it, correct it and find their way through it.
That meant the design problem was not simply:
How do we generate a transcript faster?
It was:
What should happen after the transcript exists?
1.
AI output is only the beginning
Working on Transcription Portal changed how I think about AI-assisted products. Generating useful output is only one part of the experience. The product still needs to help people understand, verify, correct and work with what the model produces. The real design problem often begins after the AI finishes its job.
2.
Give people control over generated content
Designing around transcription errors made the role of human control more concrete for me. AI can accelerate the first version of a piece of content, but users still need to decide what is correct. Making editing, correction and navigation part of the core workflow was as important as the generation itself.
3.
Design the workflow around the output
The biggest shift was moving from thinking about transcription as a task to thinking about it as a workflow. Once audio became text, new problems appeared around search, context, editing and navigation. The interface was only the visible layer. The stronger design decisions came from understanding what users needed to do with the generated content next.
